Claims decisions, on this device.
Two small open models route claims, flag legal threats and vulnerable customers, and pseudonymise letters inside this browser tab. Nothing you type leaves the device. Every answer carries its probability, so the handler decides what goes straight through.
Pseudonymise a document. Offline.
A claim letter with invented names, phone, email, IBAN, card number and address.
Drop the file
Read by this tab. The request counter below will not move.
Placeholder map (kept in this tab's memory only)
What just happened, step by step
Unzipped and read in memory. No request.
Emails, phones, IBANs, cards, dates, addresses.
"Is this a person?" with a probability, on your processor.
This looks like a prototype or a mockup. It is not: the models really downloaded into this tab, really ran on your processor, and the page works with the network switched off. Try it.
Route a claim to the right team.
Order the queue by urgency.
Flag a legal threat or a vulnerable customer.
Decisions an insurer makes all day
Run an example, then move the threshold. Above it the decision goes straight through; below it, a person looks.
Is the probability worth acting on?
Three decisions, 30 labelled messages. See how often each stated probability was right.
See the 30 messages and their labels
Every model call, measured
What a gateway like LiteLLM logs per API call: tokens, latency, cost. Kept in this tab.
Fake versions of real documents
Same wording, fake details. Repeatable with a seed.
What a pilot has to answer first
Short answers, each one something you can check on this page or in the lab's files.
Test the lock
This page carries a browser security policy that lets it connect to the site that served it and nowhere else. Try to send the text below to another website: your browser refuses before anything leaves.
Where does the text go?
Nowhere. The runtime, the tokenizer and both models run in this tab, on this device's processor. After the one-time download the request counter in the bar below stays at zero, and the browser itself forbids connections to any other site: test it on the left. DevTools → Network shows the same.
Is a pseudonymised letter anonymous?
No. Swapping names, IBANs and phone numbers for placeholders is pseudonymisation, and the result is still personal data under the GDPR (Recital 26). It reduces what reaches a partner; it does not replace a legal basis for sharing.
The redactor also misses things. Measured on 900 test documents with the rules alone: on French claim letters like the sample it found 95 % of names, and 12 % of letters still kept an identifier; on a harder public set it found far fewer. A person reads the result before it is shared.
Who makes the decision?
The handler. Every answer carries a probability and the business sets the threshold: above it a decision can go straight through, below it a person looks. Lower the threshold and more cases reach a person. A production use would go through the group's AI governance and model-risk review; this page is a prototype on synthetic data.
How accurate is it?
Measured on messages written for this page, in English, not AXA data. Claim routing: 79 % right on 29 messages; at p ≥ 0.60, 76 % go straight through and 86 % of those are right. Legal-threat and vulnerable-customer flags: AUC 0.98 on 30 messages. Urgency: rank correlation 0.44, too weak to use. Run the test above to measure it on this device.
What does a decision cost?
Nothing per call: no provider and no per-token bill. A decision takes tens of milliseconds on a laptop processor; the usage meter shows this device's own times. The costs are the one-time 59 MB download and the device's electricity.
What does a laptop need?
A recent Chrome, Edge or Firefox with WebAssembly SIMD (the device check in the bar below says whether this one qualifies), about 60 MB of browser storage and about 150 MB of memory while it runs. No installation, no admin rights, no plug-in, no GPU. The page can be installed as an app and then opens with no connection.
Which models, and under which licences?
xtremedistil-l6-h256 zero-shot (MIT, Moritz Laurer) and all-MiniLM-L6-v2 (Apache-2.0, sentence-transformers), unchanged from their public releases, run by ONNX Runtime Web (MIT). No non-commercial licence anywhere on the page. The table at the end lists them with their sources.
What about French and German?
Both models are English. French and German messages work less well, and that has not been measured here. The redactor's rules do cover French formats: NIR, IBAN, number plates, postcodes and Nom: / Prénom: labels.
What would a pilot measure?
One decision, claim routing, on a labelled sample of real claims processed in the EU: accuracy and straight-through rate against the classifier already in place, at a threshold the business chooses. The calibration test above is that experiment in miniature.
Jev, a model that answers software with decisions
Launched 15 September 2026. Typed questions in, typed answers out.
Read the full brief and our view
The company
Founded by Diogo Almeida, a former OpenAI researcher who worked on the methods behind ChatGPT, with Erik Gafni and Sasha Sheng. A $40m seed led by DCVC, at a reported $200m valuation.
What Jev returns
- Choice: one of up to 255 labelled options, with a probability for each.
- Score: a place on an ordered scale of 2–10 levels, with the distribution.
- Yes or no (Jev calls it "noul"): a probability from 0 to 1.
Training for calibration is the pitch: answers given 90 % should be right about 90 % of the time. The docs publish no calibration measurement.
The claims, and whose tests they are
- 70–500 ms end to end; 40–200× faster than frontier LLMs on "System One" tasks.
- One workflow: 193.6× faster and 444.6× cheaper than a GPT-class model. TypeSafe's own team wrote the workflows and says the gains sit at the high end.
- $0.042 per million input tokens; output is free.
- Writing on LinkedIn, Alexandre Dos Santos notes that fine-tuned classifiers already in production run 5–10× faster than Jev, at almost no cost, on in-house servers.
How you would use it, and where the data goes
- A hosted API (
POST api.typesafe.ai/v1/systemone) with Python and JavaScript SDKs. No weights, no self-hosting, no browser version. - The playground needs a sign-in (Google or an emailed code).
- The docs name no EU region. TypeSafe says it does not train on customer data, offers a data processing agreement, and offers zero data retention to enterprise customers.
The real edge is that Jev needs no task-specific training, so a new decision could go live without labelled data. For insurers the calibrated probability matters more than the speed: it sets the straight-through threshold. Act above it, refer to a person below it.
This page shows the other half: typed decisions with a probability are no longer rare. Two free models of 13 MB and 23 MB do a useful part of it on a laptop, offline. What is still scarce is a probability you can act on without labels. The telling test is one claims-routing decision, scored against the classifier already in place, with the data processed in the EU.
The open models behind this page, and others like them
Free, open, and unchanged from their public releases.
| Model | What it does here | Licence | Size (int8) | On this page |
|---|---|---|---|---|
| MoritzLaurer/xtremedistil-l6-h256-zeroshot-v1.1-all-33 | Yes / no and score: reads the message and a statement, returns the probability that the statement is true | MIT | 13.1 MB | Runs here |
| sentence-transformers/all-MiniLM-L6-v2 (ONNX by Xenova) | Choice: compares the message with short descriptions of each team | Apache-2.0 | 23.0 MB | Runs here |
| MoritzLaurer/deberta-v3-xsmall-zeroshot-v1.1-all-33 | The same zero-shot decisions, larger and more accurate | MIT | 87.2 MB | Too large for this page |
| cross-encoder/nli-deberta-v3-xsmall | Yes / no by natural-language inference | Apache-2.0 | 87.4 MB | Too large for this page |
| cross-encoder/nli-MiniLM2-L6-H768 | Yes / no by natural-language inference | Apache-2.0 | 82.8 MB | Too large for this page |
| knowledgator/gliclass-edge-v3.0 | Zero-shot classification over many labels at once | Apache-2.0 | no ONNX export | Server or Python |
Sizes and licences read from each model's Hugging Face page on 24 September 2026. The runtime is ONNX Runtime Web 1.17.3 (MIT), compiled to WebAssembly. Word files are read by mammoth.js (BSD-2-Clause). The interface follows the patterns of LocalMode (MIT).